摘要
真实语音情感识别是使人机交互更加友好的重要手段,但是训练数据稀缺为这一领域带来很多挑战。为了减小这一阻碍,提出了语句串接与重采样(UCR)方法,以便高效利用存在的训练数据。UCR方法是将原始音频样本按照情感类型进行串接,形成一个长的音频流,以一个固定粒度对其随机乱序,然后将其切割,并通过多次重采样操作来增加支持向量机(SVM)的训练样本数。实验基于一个从访谈节目中录制的真实语音情感库。实验结果表明,在统一背景模型-高斯混合模型-支持向量机(UBM—GMM—SVM)识别框架中这种训练集重构的方法错误率降低近33.10%。
Real emotion recognition can be an important means to make human-computer interaction more friendly,yet insufficient training data pose many challenges for this speech-related field.In this paper,a method to help reduce this barrier is proposed by effectively utilizing existing training data—namely,utterance concatenation and resampling（UCR）.It involves concatenation of audio files of the same emotion into a long stream,and then segmenting the stream;randomly permuting chunks of that stream;and even increasing the number of all supervectors for SVM by resampling several times.Experiments are made based on the interview speech emotion database,recorded from actual television interviews.Evaluation results show that the error rate reduction can reach 33.10% by restructuring the training data of UBM-GMM-SVM systems.
出处
《北京信息科技大学学报（自然科学版）》
2012年第2期63-67,共5页
Journal of Beijing Information Science and Technology University
基金
北京市属市管高等学校人才强教计划资助项目(PHR201007131)
